[fio.git] / workqueue.c

/*
 * Generic workqueue offload mechanism
 *
 * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
 *
 */
#include <unistd.h>

#include "fio.h"
#include "flist.h"
#include "workqueue.h"
#include "smalloc.h"
#include "pshared.h"

enum {
	SW_F_IDLE	= 1 << 0,
	SW_F_RUNNING	= 1 << 1,
	SW_F_EXIT	= 1 << 2,
	SW_F_ACCOUNTED	= 1 << 3,
	SW_F_ERROR	= 1 << 4,
};

static struct submit_worker *__get_submit_worker(struct workqueue *wq,
						 unsigned int start,
						 unsigned int end,
						 struct submit_worker **best)
{
	struct submit_worker *sw = NULL;

	while (start <= end) {
		sw = &wq->workers[start];
		if (sw->flags & SW_F_IDLE)
			return sw;
		if (!(*best) || sw->seq < (*best)->seq)
			*best = sw;
		start++;
	}

	return NULL;
}

static struct submit_worker *get_submit_worker(struct workqueue *wq)
{
	unsigned int next = wq->next_free_worker;
	struct submit_worker *sw, *best = NULL;

	assert(next < wq->max_workers);

	sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
	if (!sw && next)
		sw = __get_submit_worker(wq, 0, next - 1, &best);

	/*
	 * No truly idle found, use best match
	 */
	if (!sw)
		sw = best;

	if (sw->index == wq->next_free_worker) {
		if (sw->index + 1 < wq->max_workers)
			wq->next_free_worker = sw->index + 1;
		else
			wq->next_free_worker = 0;
	}

	return sw;
}

static bool all_sw_idle(struct workqueue *wq)
{
	int i;

	for (i = 0; i < wq->max_workers; i++) {
		struct submit_worker *sw = &wq->workers[i];

		if (!(sw->flags & SW_F_IDLE))
			return false;
	}

	return true;
}

/*
 * Must be serialized wrt workqueue_enqueue() by caller
 */
void workqueue_flush(struct workqueue *wq)
{
	wq->wake_idle = 1;

	while (!all_sw_idle(wq)) {
		pthread_mutex_lock(&wq->flush_lock);
		pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
		pthread_mutex_unlock(&wq->flush_lock);
	}

	wq->wake_idle = 0;
}

/*
 * Must be serialized by caller.
 */
void workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work)
{
	struct submit_worker *sw;

	sw = get_submit_worker(wq);
	assert(sw);

	pthread_mutex_lock(&sw->lock);
	flist_add_tail(&work->list, &sw->work_list);
	sw->seq = ++wq->work_seq;
	sw->flags &= ~SW_F_IDLE;

	pthread_cond_signal(&sw->cond);
	pthread_mutex_unlock(&sw->lock);
}

static void handle_list(struct submit_worker *sw, struct flist_head *list)
{
	struct workqueue *wq = sw->wq;
	struct workqueue_work *work;

	while (!flist_empty(list)) {
		work = flist_first_entry(list, struct workqueue_work, list);
		flist_del_init(&work->list);
		wq->ops.fn(sw, work);
	}
}

static void *worker_thread(void *data)
{
	struct submit_worker *sw = data;
	struct workqueue *wq = sw->wq;
	unsigned int ret = 0;
	FLIST_HEAD(local_list);

	sk_out_assign(sw->sk_out);

	if (wq->ops.nice) {
		if (nice(wq->ops.nice) < 0) {
			log_err("workqueue: nice %s\n", strerror(errno));
			ret = 1;
		}
	}

	if (!ret)
		ret = workqueue_init_worker(sw);

	pthread_mutex_lock(&sw->lock);
	sw->flags |= SW_F_RUNNING;
	if (ret)
		sw->flags |= SW_F_ERROR;
	pthread_mutex_unlock(&sw->lock);

	pthread_mutex_lock(&wq->flush_lock);
	pthread_cond_signal(&wq->flush_cond);
	pthread_mutex_unlock(&wq->flush_lock);

	if (sw->flags & SW_F_ERROR)
		goto done;

	while (1) {
		pthread_mutex_lock(&sw->lock);

		if (flist_empty(&sw->work_list)) {
			if (sw->flags & SW_F_EXIT) {
				pthread_mutex_unlock(&sw->lock);
				break;
			}

			if (workqueue_pre_sleep_check(sw)) {
				pthread_mutex_unlock(&sw->lock);
				workqueue_pre_sleep(sw);
				pthread_mutex_lock(&sw->lock);
			}

			/*
			 * We dropped and reaquired the lock, check
			 * state again.
			 */
			if (!flist_empty(&sw->work_list))
				goto handle_work;

			if (sw->flags & SW_F_EXIT) {
				pthread_mutex_unlock(&sw->lock);
				break;
			} else if (!(sw->flags & SW_F_IDLE)) {
				sw->flags |= SW_F_IDLE;
				wq->next_free_worker = sw->index;
				if (wq->wake_idle)
					pthread_cond_signal(&wq->flush_cond);
			}

			pthread_cond_wait(&sw->cond, &sw->lock);
		} else {
handle_work:
			flist_splice_init(&sw->work_list, &local_list);
		}
		pthread_mutex_unlock(&sw->lock);
		handle_list(sw, &local_list);
		if (wq->ops.update_acct_fn)
			wq->ops.update_acct_fn(sw);
	}

done:
	sk_out_drop();
	return NULL;
}

static void free_worker(struct submit_worker *sw, unsigned int *sum_cnt)
{
	struct workqueue *wq = sw->wq;

	workqueue_exit_worker(sw, sum_cnt);

	pthread_cond_destroy(&sw->cond);
	pthread_mutex_destroy(&sw->lock);

	if (wq->ops.free_worker_fn)
		wq->ops.free_worker_fn(sw);
}

static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
{
	pthread_join(sw->thread, NULL);
	free_worker(sw, sum_cnt);
}

void workqueue_exit(struct workqueue *wq)
{
	unsigned int shutdown, sum_cnt = 0;
	struct submit_worker *sw;
	int i;

	if (!wq->workers)
		return;

	for (i = 0; i < wq->max_workers; i++) {
		sw = &wq->workers[i];

		pthread_mutex_lock(&sw->lock);
		sw->flags |= SW_F_EXIT;
		pthread_cond_signal(&sw->cond);
		pthread_mutex_unlock(&sw->lock);
	}

	do {
		shutdown = 0;
		for (i = 0; i < wq->max_workers; i++) {
			sw = &wq->workers[i];
			if (sw->flags & SW_F_ACCOUNTED)
				continue;
			pthread_mutex_lock(&sw->lock);
			sw->flags |= SW_F_ACCOUNTED;
			pthread_mutex_unlock(&sw->lock);
			shutdown_worker(sw, &sum_cnt);
			shutdown++;
		}
	} while (shutdown && shutdown != wq->max_workers);

	sfree(wq->workers);
	wq->workers = NULL;
	pthread_mutex_destroy(&wq->flush_lock);
	pthread_cond_destroy(&wq->flush_cond);
	pthread_mutex_destroy(&wq->stat_lock);
}

static int start_worker(struct workqueue *wq, unsigned int index,
			struct sk_out *sk_out)
{
	struct submit_worker *sw = &wq->workers[index];
	int ret;

	INIT_FLIST_HEAD(&sw->work_list);

	ret = mutex_cond_init_pshared(&sw->lock, &sw->cond);
	if (ret)
		return ret;

	sw->wq = wq;
	sw->index = index;
	sw->sk_out = sk_out;

	if (wq->ops.alloc_worker_fn) {
		ret = wq->ops.alloc_worker_fn(sw);
		if (ret)
			return ret;
	}

	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
	if (!ret) {
		pthread_mutex_lock(&sw->lock);
		sw->flags = SW_F_IDLE;
		pthread_mutex_unlock(&sw->lock);
		return 0;
	}

	free_worker(sw, NULL);
	return 1;
}

int workqueue_init(struct thread_data *td, struct workqueue *wq,
		   struct workqueue_ops *ops, unsigned int max_workers,
		   struct sk_out *sk_out)
{
	unsigned int running;
	int i, error;
	int ret;

	wq->max_workers = max_workers;
	wq->td = td;
	wq->ops = *ops;
	wq->work_seq = 0;
	wq->next_free_worker = 0;

	ret = mutex_cond_init_pshared(&wq->flush_lock, &wq->flush_cond);
	if (ret)
		goto err;
	ret = mutex_init_pshared(&wq->stat_lock);
	if (ret)
		goto err;

	wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker));
	if (!wq->workers)
		goto err;

	for (i = 0; i < wq->max_workers; i++)
		if (start_worker(wq, i, sk_out))
			break;

	wq->max_workers = i;
	if (!wq->max_workers)
		goto err;

	/*
	 * Wait for them all to be started and initialized
	 */
	error = 0;
	do {
		struct submit_worker *sw;

		running = 0;
		pthread_mutex_lock(&wq->flush_lock);
		for (i = 0; i < wq->max_workers; i++) {
			sw = &wq->workers[i];
			pthread_mutex_lock(&sw->lock);
			if (sw->flags & SW_F_RUNNING)
				running++;
			if (sw->flags & SW_F_ERROR)
				error++;
			pthread_mutex_unlock(&sw->lock);
		}

		if (error || running == wq->max_workers) {
			pthread_mutex_unlock(&wq->flush_lock);
			break;
		}

		pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
		pthread_mutex_unlock(&wq->flush_lock);
	} while (1);

	if (!error)
		return 0;

err:
	log_err("Can't create rate workqueue\n");
	td_verror(td, ESRCH, "workqueue_init");
	workqueue_exit(wq);
	return 1;
}
Commit	Line	Data
a9da8ab2	1	/*
51575029	2	* Generic workqueue offload mechanism
a9da8ab2 JA	3	*
	4	* Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
	5	*
	6	*/
	7	#include <unistd.h>
	8
	9	#include "fio.h"
a9da8ab2 JA	10	#include "flist.h"
a9da8ab2 JA	11	#include "workqueue.h"
a47591e4	12	#include "smalloc.h"
ae626d4e	13	#include "pshared.h"
a9da8ab2	14
a9da8ab2 JA	15	enum {
	16	SW_F_IDLE = 1 << 0,
	17	SW_F_RUNNING = 1 << 1,
	18	SW_F_EXIT = 1 << 2,
d22042d2 JA	19	SW_F_ACCOUNTED = 1 << 3,
d22042d2 JA	20	SW_F_ERROR = 1 << 4,
a9da8ab2 JA	21	};
	22
	23	static struct submit_worker __get_submit_worker(struct workqueue wq,
	24	unsigned int start,
	25	unsigned int end,
	26	struct submit_worker **best)
	27	{
	28	struct submit_worker *sw = NULL;
	29
	30	while (start <= end) {
	31	sw = &wq->workers[start];
	32	if (sw->flags & SW_F_IDLE)
	33	return sw;
	34	if (!(best) \|\| sw->seq < (best)->seq)
	35	*best = sw;
	36	start++;
	37	}
	38
	39	return NULL;
	40	}
	41
	42	static struct submit_worker get_submit_worker(struct workqueue wq)
	43	{
	44	unsigned int next = wq->next_free_worker;
	45	struct submit_worker sw, best = NULL;
	46
	47	assert(next < wq->max_workers);
	48
	49	sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
	50	if (!sw && next)
	51	sw = __get_submit_worker(wq, 0, next - 1, &best);
	52
	53	/*
	54	* No truly idle found, use best match
	55	*/
	56	if (!sw)
	57	sw = best;
	58
	59	if (sw->index == wq->next_free_worker) {
	60	if (sw->index + 1 < wq->max_workers)
	61	wq->next_free_worker = sw->index + 1;
	62	else
	63	wq->next_free_worker = 0;
	64	}
	65
	66	return sw;
	67	}
	68
1391052a	69	static bool all_sw_idle(struct workqueue *wq)
a9da8ab2 JA	70	{
	71	int i;
	72
	73	for (i = 0; i < wq->max_workers; i++) {
	74	struct submit_worker *sw = &wq->workers[i];
	75
	76	if (!(sw->flags & SW_F_IDLE))
1391052a	77	return false;
a9da8ab2 JA	78	}
a9da8ab2 JA	79
1391052a	80	return true;
a9da8ab2 JA	81	}
	82
	83	/*
	84	* Must be serialized wrt workqueue_enqueue() by caller
	85	*/
	86	void workqueue_flush(struct workqueue *wq)
	87	{
	88	wq->wake_idle = 1;
	89
	90	while (!all_sw_idle(wq)) {
	91	pthread_mutex_lock(&wq->flush_lock);
	92	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	93	pthread_mutex_unlock(&wq->flush_lock);
	94	}
	95
	96	wq->wake_idle = 0;
	97	}
	98
	99	/*
0943e13c	100	* Must be serialized by caller.
a9da8ab2	101	*/
26de50cf	102	void workqueue_enqueue(struct workqueue wq, struct workqueue_work work)
a9da8ab2 JA	103	{
	104	struct submit_worker *sw;
	105
	106	sw = get_submit_worker(wq);
26de50cf JA	107	assert(sw);
	108
	109	pthread_mutex_lock(&sw->lock);
	110	flist_add_tail(&work->list, &sw->work_list);
	111	sw->seq = ++wq->work_seq;
	112	sw->flags &= ~SW_F_IDLE;
a9da8ab2	113
26de50cf	114	pthread_cond_signal(&sw->cond);
b93a8eb0	115	pthread_mutex_unlock(&sw->lock);
a9da8ab2 JA	116	}
	117
	118	static void handle_list(struct submit_worker sw, struct flist_head list)
	119	{
	120	struct workqueue *wq = sw->wq;
88271841	121	struct workqueue_work *work;
a9da8ab2 JA	122
a9da8ab2 JA	123	while (!flist_empty(list)) {
88271841 JA	124	work = flist_first_entry(list, struct workqueue_work, list);
88271841 JA	125	flist_del_init(&work->list);
ee2b6d6e	126	wq->ops.fn(sw, work);
a9da8ab2 JA	127	}
	128	}
	129
a9da8ab2 JA	130	static void worker_thread(void data)
	131	{
	132	struct submit_worker *sw = data;
	133	struct workqueue *wq = sw->wq;
d22042d2	134	unsigned int ret = 0;
a9da8ab2 JA	135	FLIST_HEAD(local_list);
a9da8ab2 JA	136
24660963 JA	137	sk_out_assign(sw->sk_out);
24660963 JA	138
f6496ba7 JA	139	if (wq->ops.nice) {
	140	if (nice(wq->ops.nice) < 0) {
	141	log_err("workqueue: nice %s\n", strerror(errno));
	142	ret = 1;
	143	}
	144	}
	145
	146	if (!ret)
	147	ret = workqueue_init_worker(sw);
	148
a9da8ab2 JA	149	pthread_mutex_lock(&sw->lock);
	150	sw->flags \|= SW_F_RUNNING;
	151	if (ret)
	152	sw->flags \|= SW_F_ERROR;
	153	pthread_mutex_unlock(&sw->lock);
	154
	155	pthread_mutex_lock(&wq->flush_lock);
	156	pthread_cond_signal(&wq->flush_cond);
	157	pthread_mutex_unlock(&wq->flush_lock);
	158
	159	if (sw->flags & SW_F_ERROR)
	160	goto done;
	161
	162	while (1) {
	163	pthread_mutex_lock(&sw->lock);
	164
	165	if (flist_empty(&sw->work_list)) {
	166	if (sw->flags & SW_F_EXIT) {
	167	pthread_mutex_unlock(&sw->lock);
	168	break;
	169	}
	170
ee2b6d6e	171	if (workqueue_pre_sleep_check(sw)) {
a9da8ab2	172	pthread_mutex_unlock(&sw->lock);
ee2b6d6e	173	workqueue_pre_sleep(sw);
a9da8ab2 JA	174	pthread_mutex_lock(&sw->lock);
	175	}
	176
	177	/*
	178	* We dropped and reaquired the lock, check
	179	* state again.
	180	*/
	181	if (!flist_empty(&sw->work_list))
	182	goto handle_work;
	183
	184	if (sw->flags & SW_F_EXIT) {
	185	pthread_mutex_unlock(&sw->lock);
	186	break;
	187	} else if (!(sw->flags & SW_F_IDLE)) {
	188	sw->flags \|= SW_F_IDLE;
	189	wq->next_free_worker = sw->index;
	190	if (wq->wake_idle)
	191	pthread_cond_signal(&wq->flush_cond);
	192	}
17ecadc3	193
a9da8ab2 JA	194	pthread_cond_wait(&sw->cond, &sw->lock);
	195	} else {
	196	handle_work:
	197	flist_splice_init(&sw->work_list, &local_list);
	198	}
	199	pthread_mutex_unlock(&sw->lock);
	200	handle_list(sw, &local_list);
4cf30b66 VF	201	if (wq->ops.update_acct_fn)
4cf30b66 VF	202	wq->ops.update_acct_fn(sw);
a9da8ab2 JA	203	}
a9da8ab2 JA	204
a9da8ab2	205	done:
24660963	206	sk_out_drop();
a9da8ab2 JA	207	return NULL;
	208	}
	209
c35c582d	210	static void free_worker(struct submit_worker sw, unsigned int sum_cnt)
a9da8ab2	211	{
ee2b6d6e	212	struct workqueue *wq = sw->wq;
a9da8ab2	213
c35c582d	214	workqueue_exit_worker(sw, sum_cnt);
a9da8ab2 JA	215
	216	pthread_cond_destroy(&sw->cond);
	217	pthread_mutex_destroy(&sw->lock);
ee2b6d6e JA	218
	219	if (wq->ops.free_worker_fn)
	220	wq->ops.free_worker_fn(sw);
a9da8ab2 JA	221	}
	222
	223	static void shutdown_worker(struct submit_worker sw, unsigned int sum_cnt)
	224	{
a9da8ab2	225	pthread_join(sw->thread, NULL);
c35c582d	226	free_worker(sw, sum_cnt);
a9da8ab2 JA	227	}
	228
	229	void workqueue_exit(struct workqueue *wq)
	230	{
	231	unsigned int shutdown, sum_cnt = 0;
	232	struct submit_worker *sw;
	233	int i;
	234
f8e266c9 JA	235	if (!wq->workers)
	236	return;
	237
a9da8ab2 JA	238	for (i = 0; i < wq->max_workers; i++) {
	239	sw = &wq->workers[i];
	240
	241	pthread_mutex_lock(&sw->lock);
	242	sw->flags \|= SW_F_EXIT;
	243	pthread_cond_signal(&sw->cond);
	244	pthread_mutex_unlock(&sw->lock);
	245	}
	246
	247	do {
	248	shutdown = 0;
	249	for (i = 0; i < wq->max_workers; i++) {
	250	sw = &wq->workers[i];
	251	if (sw->flags & SW_F_ACCOUNTED)
	252	continue;
b7d0bbf1	253	pthread_mutex_lock(&sw->lock);
a9da8ab2	254	sw->flags \|= SW_F_ACCOUNTED;
b7d0bbf1	255	pthread_mutex_unlock(&sw->lock);
a9da8ab2 JA	256	shutdown_worker(sw, &sum_cnt);
	257	shutdown++;
	258	}
	259	} while (shutdown && shutdown != wq->max_workers);
	260
a47591e4	261	sfree(wq->workers);
f8e266c9	262	wq->workers = NULL;
a9da8ab2 JA	263	pthread_mutex_destroy(&wq->flush_lock);
a9da8ab2 JA	264	pthread_cond_destroy(&wq->flush_cond);
2a274336	265	pthread_mutex_destroy(&wq->stat_lock);
a9da8ab2 JA	266	}
a9da8ab2 JA	267
24660963 JA	268	static int start_worker(struct workqueue *wq, unsigned int index,
24660963 JA	269	struct sk_out *sk_out)
a9da8ab2 JA	270	{
	271	struct submit_worker *sw = &wq->workers[index];
	272	int ret;
	273
	274	INIT_FLIST_HEAD(&sw->work_list);
34febb23 JA	275
34febb23 JA	276	ret = mutex_cond_init_pshared(&sw->lock, &sw->cond);
f9e5b5ee JK	277	if (ret)
f9e5b5ee JK	278	return ret;
34febb23	279
a9da8ab2 JA	280	sw->wq = wq;
a9da8ab2 JA	281	sw->index = index;
24660963	282	sw->sk_out = sk_out;
a9da8ab2	283
ee2b6d6e JA	284	if (wq->ops.alloc_worker_fn) {
	285	ret = wq->ops.alloc_worker_fn(sw);
	286	if (ret)
	287	return ret;
	288	}
	289
a9da8ab2 JA	290	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
	291	if (!ret) {
	292	pthread_mutex_lock(&sw->lock);
	293	sw->flags = SW_F_IDLE;
	294	pthread_mutex_unlock(&sw->lock);
	295	return 0;
	296	}
	297
c35c582d	298	free_worker(sw, NULL);
a9da8ab2 JA	299	return 1;
	300	}
	301
	302	int workqueue_init(struct thread_data td, struct workqueue wq,
24660963 JA	303	struct workqueue_ops *ops, unsigned int max_workers,
24660963 JA	304	struct sk_out *sk_out)
a9da8ab2 JA	305	{
	306	unsigned int running;
	307	int i, error;
f9e5b5ee	308	int ret;
a9da8ab2	309
24660963	310	wq->max_workers = max_workers;
a9da8ab2	311	wq->td = td;
5bb79f69	312	wq->ops = *ops;
a9da8ab2 JA	313	wq->work_seq = 0;
a9da8ab2 JA	314	wq->next_free_worker = 0;
f9e5b5ee	315
34febb23 JA	316	ret = mutex_cond_init_pshared(&wq->flush_lock, &wq->flush_cond);
34febb23 JA	317	if (ret)
f9e5b5ee	318	goto err;
34febb23 JA	319	ret = mutex_init_pshared(&wq->stat_lock);
34febb23 JA	320	if (ret)
f9e5b5ee	321	goto err;
a9da8ab2	322
a47591e4	323	wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker));
81b3c86f JA	324	if (!wq->workers)
81b3c86f JA	325	goto err;
a9da8ab2 JA	326
a9da8ab2 JA	327	for (i = 0; i < wq->max_workers; i++)
24660963	328	if (start_worker(wq, i, sk_out))
a9da8ab2 JA	329	break;
	330
	331	wq->max_workers = i;
bddc8d16 JA	332	if (!wq->max_workers)
bddc8d16 JA	333	goto err;
a9da8ab2 JA	334
	335	/*
	336	* Wait for them all to be started and initialized
	337	*/
	338	error = 0;
	339	do {
	340	struct submit_worker *sw;
	341
	342	running = 0;
	343	pthread_mutex_lock(&wq->flush_lock);
	344	for (i = 0; i < wq->max_workers; i++) {
	345	sw = &wq->workers[i];
	346	pthread_mutex_lock(&sw->lock);
	347	if (sw->flags & SW_F_RUNNING)
	348	running++;
	349	if (sw->flags & SW_F_ERROR)
	350	error++;
	351	pthread_mutex_unlock(&sw->lock);
	352	}
	353
	354	if (error \|\| running == wq->max_workers) {
	355	pthread_mutex_unlock(&wq->flush_lock);
	356	break;
	357	}
	358
	359	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	360	pthread_mutex_unlock(&wq->flush_lock);
	361	} while (1);
	362
bddc8d16 JA	363	if (!error)
bddc8d16 JA	364	return 0;
a9da8ab2	365
bddc8d16 JA	366	err:
	367	log_err("Can't create rate workqueue\n");
	368	td_verror(td, ESRCH, "workqueue_init");
	369	workqueue_exit(wq);
	370	return 1;
a9da8ab2	371	}